Disclosed are systems, methods, and products for language learning that automatically extracts keywords from resources using various natural-language processing product features, which can be combined with custom-designed learning activities to offer a needs-based, adaptive learning methodology. The system may receive resources having text and then determine a text difficulty score that predicts how difficult the resource is for language learners based on any number of factors, including any number of semantic and syntactic features of the text. Training resources labeled with metadata may be used to train a statistical model for determining difficulty scores of newly received text. Resources may be grouped based on difficulty score, and groups of resources may correspond to language learners' proficiency levels.